h-infinity control based scheduler for the deployment of ... · h-infinity control based...
TRANSCRIPT
H-Infinity Control Based Scheduler for the Deployment
of Small Cell Networks
Subhash Lakshminarayana,1,2,∗, Mohamad Assaad1,∗, Merouane Debbah2,∗
Abstract
In this work, we address the joint problem of traffic scheduling and interference
management related to the deployment of Small Cell Networks (SCNs). The
base stations of the SCNs (which we will refer to as Micro Base Stations, MBSs)
are low power devices with limited buffer size. They are connected to a Central
Scheduler (CS) with limited capacity backhaul links. In this scenario, traffic
has to be scheduled from the network to the MBS queues in such a way that the
queue-length at MBS remains as close as possible to a given target queue-length.
The challenge is to design a scheduler which is oblivious to the wireless link be-
tween the MBSs and the User Terminals (UTs). For the traffic arriving at the
MBS, we need to efficiently transmit it over the wireless channel to the UTs in
an interference limited environment. Additionally, real time centralized inter-
ference management techniques will not be feasible. In this paper, we decouple
the joint scheduling and interference management into two separate parts. For
the scheduling problem, we propose a H∞ control based scheduler which regu-
lates the arrival rates to the queues at the MBS. For the problem of interference
management over the wireless channel, we propose a multi-cell beamforming
technique and formulate a decentralized algorithm using tools from the field
∗Corresponding author.Email addresses: [email protected] (Subhash Lakshminarayana,),
[email protected] (Mohamad Assaad), [email protected] (MerouaneDebbah)
1Department of Telecommunications, SUPELEC, 3 rue Joliot-Curie, 91192 Gif-sur-Yvette,France
2Alcatel-Lucent Chair on Flexible Radio, SUPELEC, 3 rue Joliot-Curie, 91192 Gif-sur-Yvette, France
Preprint submitted to Performance Evaluation September 2, 2011
of random matrix theory. The beamforming vectors are designed to optimize
two performance metrics of interest namely downlink power minimization and
weighted sum rate maximization. Our simulation results show that the H∞
based queue length control algorithm stabilizes the queue-lengths at the MBS
and keeps the variation of the queue-length around the target to a minimum.
Keywords: Small cell netwoks, Random matrix theory, Scheduler, Beam-
forming, H∞ control
1. Introduction
According to the observation made by Martin Cooper of Arraycomm, during
the last century, the capacity of wireless networks has doubled every 30 months.
It has been observed that the biggest gains in the efficiency of spectrum uti-
lization has been due to network densification, i.e., shrinking of cell sizes [1].
Hence the next paradigm shift in the field of cellular networks has been identi-
fied as shrinkage of cell sizes resulting in Small Cells Networks, (SCN) [2]. The
reduction in the cell sizes implies reducing the distance between the transmit-
ter and the receiver, creating the dual benefit of higher link qualities and more
spatial reuse. It also implies reduction in the amount of power transmitted. In
this document, we will refer to the SCN base stations as Micro Base Stations
(MBSs). Typically the MBSs of the small cells are of the form of data access
points with limited battery and buffer sizes installed by the users to provide
better data and voice coverage over a small area.
However, the reduction of cell sizes imposes numerous technical challenges
in the design of such SCNs. Some of the technical issues faced in the implemen-
tation of SCNs are the problem of interference management, traffic scheduling,
optimal cell size planning, call handover issues, security, backhaul infrastruc-
ture etc. For a detailed survey on the SCN deployment and design challenges
in SCNs, please refer to [2],[3],[4].
In this work, we address two of the issues associated with the deployment
SCNs namely interference management and traffic flow regulation.
2
CS
Q1
Q2
Q3
UT1
UT2
UT3
MBS1
MBS2
MBS3
Figure 1: System Model
Consider the scenario in Figure 1. The set up consists of small cells with
MBS serving the UTs over the wireless channel. The MBSs are equipped with
multiple antennas and the UTs have a single antenna each. Inter-cell interfer-
ence is a major bottleneck in achieving good data rates for the UTs, specially
the UTs at the cell edge. Hence the MBSs perform joint multi-cell beamforming
in order to manage interference. However, the MBSs must perform beamform-
ing in a decentralized manner. To this end, we use joint multi-cell beamforming
technique developed in [5] using tools from random matrix theory [6]. We con-
sider two performance metrics of interest namely downlink power minimization
and weighted sum rate maximization.
The MBSs are in turn connected to a Central Station (CS) which forwards
the traffic arriving from the backbone network to the MBSs through wired
backhaul link. We assume that the CS is an infinite reservoir of packets. Since
the buffer sizes at the MBS is limited, we need an efficient flow controller which
regulates the traffic from the CS to the MBS in such a way that the queue-length
at the MBS remains as close as possible to a given target queue-length (usually
the capacity of the buffer at the MBS). The main challenge in the design of the
3
flow controller is that it must regulate the traffic flow while being oblivious to
the wireless channel conditions between the MBS and the UTs. This is a very
practical assumption, since the CS is typically connected to a large number
of MBSs and hence the cannot keep track of the channel conditions of all the
MBSs.
We model the queue-length evolution as a linear dynamic system and de-
velop a robust queue-length controller/regulator based on H∞ control [7]. The
H∞ type robust controller controls the trajectory of a linear dynamical system
under unknown disturbances (noise) without the knowledge of the statistical
distribution of the noise process. The advantage of such a control algorithm in
our setting is that the MBS do not need to feedback the CSI of the UTs to the
CS. The task of the H∞ controller is to specify an instantaneous arrival rate
every time slot (from the CS to the queues present at the MBS) such that it
minimizes the variation of queue length size around the target queue length.
1.1. Comparison with Cross-Layer Optimization in Wireless Networks
The problem handled in this paper is multi-disciplinary and spans the fields
scheduling and power allocation in wireless networks and decentralized interfer-
ence management techniques in the context of cellular networks. The problem
set up described above has a flavor of cross-layer design to it. Cross-layer opti-
mization has been a well studied topic in the context of single-hop and multi-hop
wireless networks [8], [9], [10]. In these works, the problem framework naturally
leads to decomposition of the problem into two parts. A fairness based flow
controller to regulate the traffic flow into the queues and a max-weight based
scheduler to select a set of flows whose packets can be scheduled for transmission
at each time slot.
Our work is loosely connected to the cross layer framework of the above
works. In our work, the scheduling aspect of cross-layer design is replaced
by beamforming on the wireless channel. Equipped with multiple antennas,
the MBSs obtain spatial degrees of freedom for interference management. The
MBSs can simultaneously serve multiple UTs by beamforming on the wireless
4
channel. The flow controller part is replaced by an H∞ based flow controller.
However, unlike the fairness based congestion controllers in previous works, the
objective of the H∞ based flow controller is to minimize the variation of the
queue-length around a given target queue-length. The application ofH∞ control
for queue-length stabilization in the context of wireless network has been used in
[11],[12]. In particular, [12] addresses a dynamic resource allocation problem in
multi-service downlink OFDMA systems. However, to the authors’ knowledge,
the joint problem of scheduling using H∞ control theory, power allocation and
beamforming has not been handled before in the context of cellular networks.
Throughout this work, we use boldface lowercase and uppercase letters to
designate column vectors and matrices, respectively. For a matrix X, Xi,j
denotes the (i, j) entry of X. XT and XH denote the transpose and complex
conjugate transpose of matrix X. We denote an identity matrix of size M as IM
and diag(x1, ..., xM ) is a diagonal matrix of size M with the elements xi on its
main diagonal. We use x ∼ CN (m,R) to state that the vector x has a complex
Gaussian distribution with mean m and covariance matrix R. We will use the
notation (x)+ = max(x, 0). We will also use the notation ||x||2W to denote the
weighted second order norm of the vector x given by xHWx. We use E [Y ] to
denote the expected value of a random variable Y.
The rest of the paper is organized as follows. In section 2, we provide the sys-
tem model and introduce the notations. We introduce the problem formuation
in two scenarios namely downlink power minimization and weighted sum rate
maximization. In section 3, we provide the algorithm description for the decen-
tralized beamforming design in the two scenarios described above. In section
4, we provide the details of the flow control algorithm based on H∞ control.
In section 5, we compare the performance of the H∞ against standard LQG
based flow control. The simulation results are provided in section 6. We also
provide a brief description of the details of H∞ based control and LQG control
in Appendix A and B respectively.
5
2. System Model
The system model is shown in Figure 1. It can broken down into two parts.
The frontend consisting of the MBSs and the UTs and the backend consisting
of the CS and the backhaul links.
We focus on the frontend first. We consider a SCN scenario consisting N
cells and K UTs per cell. The UTs in each cell are served by their MBS which
is equipped with Nt antennas and each UT has a single antenna. The nota-
tion UTi,j denotes the j-th UT present in the i-th cell. The MBS of each
cell serves only the UTs present only in its cell. Let hti,j,k ∈ CNt denote the
channel from the MBS i to the UTk,j during the coherence interval t. We as-
sume that the elements of the channel vector are Gaussian distributed, i.e.,
hti,j,k ∼ CN (0, (σ2
i,j,k/Nt)INt), 3 the variance of which depends upon the path
loss model between BS i and UTj,k. The channel variance has been scaled by
the factor Nt to maintain the per antenna power constraint at each base station.
We denote the coherence interval of the channel by the notation Tc.
We consider that a discrete time system in which the time slots are indexed
by t. Before proceeding further, we will clarify the notion of time slot t in our
context. We assume that arrivals and departures from the queues present at the
MBS happen once every coherence interval of the channel. Hence the notation
t indexes the coherence periods of the channel.
The backend part consists of the CS and the backhaul links connecting the
CS to the MBS. The task of the CS is to schedule the traffic arriving from the
underlying wired backbone into the queues present at the MBS. We assume that
the backhaul links have just enough capacity to deliver all the traffic scheduled
by the CS during a given coherence.
Let us denote the queue length at MBS i for the UTi,j in its cell during the
time slot t by the notation qti,j . We denote the target buffer length at the MBS
3Note that the variance of the channel is constant across all time slots since we consider
that the UTs do not move.
6
by the notation qi,j .We denote the traffic arrival rate into the queue of the MBS
i for UTi,j during the time slot t by the notation ati,j . We also denote the rate
at which packets are transmitted out of the queue (i, j) into the wireless link,
(i.e. the service rate) during the time slot t by the notation µti,j .
We now describe the design parameters of the problem. Let wti,j ∈ CNt de-
note the transmit downlink beamforming vector for the UTi,j during time slot t.
Likewise, let Γti,j denote the received SINR for UTi,j and γti,j the corresponding
target SINR during the time slot t. The received signal yti,j ∈ C for the UTi,j
during the time slot t is given by
yti,j =
K∑
l=1
htH
i,i,jwti,lx
ti,l +
N∑
m=1m 6=i
K∑
n=1
htH
m,i,jwtm,nx
tm,n + zti,j
where xti,j ∈ C represents the information signal for the UTi,j during the time
slot t and zti,j ∼ CN (0, σ2) is the corresponding additive white Gaussian complex
noise.
We now provide two different problem formulations depending on the metric
of interest namely the downlink power minimization and the weighted sum rate
maximization.
Downlink Power Minimization - PMIN : The downlink power minimization
and scheduler design problem can be cast into the following optimization prob-
lem given by
PMIN : minwt
i,j
∑
i,j
wtH
i,jwti,j , ∀t = 1 . . . T (1)
s.t. Γti,j ≥ γti,j , t = 1 . . . T, i = 1 . . .N, j = 1 . . .K
limT→∞
1
T
T∑
t=1
qti,j = qi,j
The objective function of the optimization problem in (1) attempts to mini-
mize the total power transmitted in the downlink during each time slot t. The
first constraint equation is the target SINR constraint for each UT. The sec-
ond constraint equation denotes the scheduler design constraint which tries to
maintain the long term average of the queue-lengths (across multiple coherence
7
intervals) as close as possible to the target queue-lengths. The received SINR
in the downlink is given by the following expression
Γti,j =
|wtHi,jh
ti,i,j |
2
∑
l 6=j |wtHi,lh
ti,i,j |
2 +∑
m 6=i,n |wtHm,nh
tm,i,j |
2 + σ2(2)
The numerator term is the useful signal. The denominator terms represent the
intra-cell interference, inter-cell interference and the thermal noise (in order as
they appear in the denominator).
Weighted Sum Rate Maximization - RMAX : We now consider the problem
of maximizing the weighted sum rate subject to max-power constraint on each
MBS. The optimization problem is given by
RMAX : maxwt
i,j
∑
i,j
βi,j log(1 + Γti,j), ∀t = 1 . . . T (3)
s.t.∑
j
wtH
i,jwti,j ≤ Pmax, t = 1 . . . T, i = 1 . . .N
limT→∞
1
T
T∑
t=1
qti,j = qi,j
where the downlink SINR Γti,j is given as in equation (2). Here, βi,j are scalar
weights for the downlink rate of UTi,j and Pmax is the max-power constraint
per MBS.
3. Algorithm Design
The basic optimization problem in (1) and (3) are difficult to solve due to
the interdependencies of the various parameters involved. Hence, in this work
we decouple the problem into two stages. The first stage is the scheduling or
the rate control problem for the queues of the MBS. The design objective of the
scheduler is to minimize the variance of the queue-length around a given target
queue-length. The second stage is the interference management problem.
The parameters connecting the two problems are the SINR and the service
rates of the queues. The relationship between the service rate of the queues at
8
the MBS, µ and the SINR, γ, is given by the Shannon-Hartley theorem 4
µ = B log(1 + γ) (4)
where B is the total number of channel uses available during the coherence
interval Tc. Accordingly, we denote the service rate corresponding to the target
SINR (γi,j) for UTi,j by µi,j and the service rate corresponding to the achieved
SINR in the downlink (Γti,j) during each time slot t by µt
i,j . We will refer to µi,j
as the target service rate for the queues.
The decoupled algorithm can be summarized as follows:
• The UTs request a target SINR during each time slot depending on their
requirement.
• The beamformer design algorithm computes the optimal design parame-
ters to serve the data to the UTs.
• During each time slot, the scheduler regulates the amount of traffic arriv-
ing into the queues of the MBS such that the queues are stable.
In the subsequent part of this section, we will describe the decentralized multi-
cell beamforming algorithm in detail. We will defer the flow regulation part to
section 4. The problem of decentralized multicell beamforming has received con-
siderable attention in recent years [13], [14], [15], [16]. [13] suggests beamform-
ing design based on maximizing the virtual signal to interference and noise ratio
(VSINR) metric which is shown to attain optimal rate points. [14] proposes a
distributed beamforming approach based on Kalman smoothing involving mes-
sage passing between the BSs. [15], [16] propose decentralized beamforming
technique based on localized message passing, hence eliminating the need for a
central processor. The decentralized algorithm proposed in this work is based
on random matrix theory based approximations. In our algorithm, the MBSs
4We assume that the coherence interval of the channel is long enough for the transmitter
to achieve the channel capacity.
9
would need to exchange only the channel statistics between themselves and not
the instantaneous CSI, hence reducing the information exchange. Such algo-
rithms are asymptotically optimal (when the system dimensions become very
large, to be explained later in this section) and provide good approximations
for practical network dimensions. We now describe the algorithm in detail.
3.1. Beamformer Design and Power Allocation: The Downlink Power Mini-
mization Case
We first provide the a distributed algorithm to compute the optimal beam-
forming vectors and downlink power minimization problem in (1). We take
up the beamformer design problem during a given time slot t. Hence we drop
the superscript t in subsequent equations related to the above problem. The
optimization problem for the beamforming design is given by
minwi,j
∑
i,j
wHi,jwi,j (5)
s.t. Γi,j ≥ γi,j , i = 1 . . .N, j = 1 . . .K
The beamformer design problem is inspired from our previous work in [5].
The downlink power minimization problem in equation (5) is more difficult
to solve since the computation of the beamforming vector of a given UT in turn
depends on the beamforming vectors of the other UTs as well. Hence, using an
approach similar to the one in [17], we convert this problem into a dual uplink
problem. The uplink problem is much easier to be solved. The dual problem in
the uplink is given as
max∑
i,j
λi,jσ2 (6)
s.t. Λi,j ≥ γi,j , i = 1, . . . , N, j = 1, . . . ,K
where the uplink SINR, Λi,j is given by
Λi,j =λi,j |wH
i,jhi,i,j |2∑
(m,l) 6=(i,j) λm,l|wHi,jhi,m,l|2 + αi||wi,j ||22
10
where w denotes the corresponding uplink beamforming vector and λi,j repre-
sents the dual variable associated with the optimization problem in (5). The
λi,j can be viewed as the dual uplink power.
Before stating the optimal beamforming algorithm, we define the follow-
ing matrices. Let Hi = [hi,1,1hi,1,2 · · ·hi,m,n · · ·hi,N,K ] be the matrix whose
columns are formed by the channel vectors from BS i to all the UTs across all
the cells. Similarly, Λ = diag[λ1,1λ1,2 · · ·λm,n · · ·λN,K ] be a diagonal matrix
with diagonal elements being the uplink power allocations. We also define the
matrix Σi as
Σi = HiΛHHi (7)
We define the Stieltjes transform ([6]) of the empirical eigen value distribution
of the matrix Σi by the notation mΣi. When the number of antennas Nt on
the MBS and the number of UTs per cell K, become very large, (Nt,K → ∞)
5
• The iterative function for evaluating the optimal uplink power allocation
λi,j is given by
λi,j =(
(
1 +1
γi,j)( σ2
i,i,jmΣi(−αi)
1 + σ2i,i,jλi,jmΣi
(−αi)
)
)−1
(8)
• The optimal receive uplink beamformers can be evaluated as
wi,j =(
∑
m,l
λm,lhi,m,lhHi,m,l + αiI
)−1
hi,i,j (9)
• The optimal transmit downlink beamformers are computed using
wi,j =√
δi,jwi,j (10)
5The details of the algorithm can be found in [5].
11
The parameter δi,j is a scaling factor between the uplink and the downlink
beamforming vectors. It can be computed as in [5].
The effectiveness of our algorithm lies in the fact that the computation of
the uplink power parameter λ depends only on the channel statistics and not
upon the actual channel realization. Hence in order to compute the beamform-
ing vectors, MBSs only need to exchange the channel statistics (of their UTs)
between themselves. This tremendously reduces the computational complexity
and the information to be exchanged between between the BSs especially in a
fast fading scenario in which the channel realization changes rapidly where as
the channel statistics remain constant over fairly long periods of time (assuming
low mobility scenario).
Note however, that the algorithm is optimal in the asymptotic setting im-
plying that the SINR constraints are perfectly satisfied only when the number
of antennas on the MBS and the number of UTs tend to infinity. However, in
a practical scenario, when the network dimensions are finite, the authors in [5]
show with the help of simulation results that the actual achieved SINR fluctu-
ates around the target SINR. The fluctuations around the target SINR depends
upon the channel realizations across the multiple coherence intervals.
3.1.1. The Problem of Feasibility
We now include a practical constraint in the beamformer design problem
namely the max-power constraint on each MBS. We rewrite the optimization
problem of Equation (5) incorporating the max-power constraint.
minwi,j
∑
i,j
wHi,jwi,j (11)
s.t. Γi,j ≥ γi,j , i = 1 . . .N, j = 1 . . .K∑
j
wHi,jwi,j ≤ Pmax i = 1 . . .N
The optimization problem in (5) has a solution only if the max-power constraints
for each MBS is satisfied. In other words, if the problem is infeasible, it implies
that given the max-power constraints on the MBSs, the targeted SINR lie out-
side the capacity region of the system. In this work, we address this problem
12
by selectively reducing the target SINR of a few UTs. We provide a heuristic
approach to select the UTs whose target SINRs can be reduced. For the cells
in which∑
j wHi,jwi,j ≥ Pmax, do the following steps.
1. Denote Pi,j as the downlink power for the UTi,j given by Pi,j = wHi,jwi,j .
Among the UTs belonging to the cells for which power constraint is vi-
olated, select the set of UT(s) which has the maximum ratio δi,j =Pi,j
qi,j.
We define the UTi,j′ = argmaxj δi,j .
2. If the target SINR γi,j′ > τ, decrease the target SINR for UTi,j′ by a
small amount ∆. Else, choose the UT(s) for which the parameter δi,j′ is
the next highest.
3. Solve the optimization problem in Equation (5) with the new set of target
SINR constraints.
4. Repeat the steps 1− 4 until the problem becomes feasible.
The intuition behind the algorithm is as follows. During each time slot t, if the
max-power constraints for the BSs is not met, we start by choosing the set of
UTs for which δi,j is maximum. The numerator term is the power consumed
by the UT. If the UT is consuming a lot of power, then the UT probably has a
bad channel realization and hence we reduce its target SINR. The denominator
term is the queue-length at time slot t. Minimum queue-length implies that
the UT(s) has lesser amount of information bits to be served and hence we
can afford to reduce the data rate of that UT. The parameter ∆ is a design
parameter which can be chosen suitably depending on the simulation scenario.
The parameter τ is a small quantity which ensures that each UT gets at least a
minimum rate. No UT goes unserved.
Therefore, the actual achieved SINR in the downlink fluctuates around the
target SINR both due to the asymptotic nature of the decentralized beamformer
design algorithm and also the feasibility issue. These fluctuations of the SINR
result in a variation of queue-length around a given target queue-length. More-
over, the fluctuations depend on the actual channel realization during each time
slot of which the CS is oblivious. Hence, we need an effective queue-length con-
13
trol algorithm which accounts for the fluctuations of the SINR. We will defer
the description of the H∞ based flow controller to Section 4.
3.2. Beamformer Design and Power Allocation: The Weighted Sum-rate Maxi-
mization Case
We now take up the weighted sum rate maximization case in (3). As in
the case of downlink power minimization problem discussed in the previous
subsection, we decompose the problem in (3) into two parts namely the sum-
rate maximization and queue-length stabilization. The problem of weighted
sum rate maximization is given by
maxwt
i,j
∑
i,j
βi,j log(1 + Γti,j), ∀t = 1 . . . T (12)
s.t.∑
j
wtH
j wti,j ≤ Pmax, t = 1 . . . T, i = 1 . . .N
Optimizing over the beamforming vector in the case of MIMO downlink channel
is a particularly challenging problem even in the centralized scenario (MIMO
broadcast channel) which has been addressed in the past literature [18],[19].
Noting that the main objective of the paper is to bring out the effectiveness
of the H-Infinity queue length controller, we avoid the complications associated
with the above optimization problem by fixing the direction of the beamforming
vector and optimize only over the power allocation associated with the UTs.
We choose the MMSE beamforming vector to fix the direction of beamforming
vector (di,j) given by
di,j =
(
∑
m,l hi,m,lhHi,m,l + I
)−1
hi,i,j
||(
∑
m,l hi,m,lhHi,m,l + αiI
)−1
hi,i,j ||2
(13)
and the beamforming vector wi,j is given by
wi,j = Pi,jdi,j
where Pi,j is the power allocation to the UTi,j . As shown in the iterative algo-
rithm in [18],[19], the MMSE beamforming direction is a reasonable choice for
14
optimizing the weighted sum rate. Hence the modified optimization problem is
given by
maxP t
i,j
∑
i,j
βi,j log
(
1 +P ti,j |d
tHi,jh
ti,i,j |
2
∑
l 6=j Pti,l|d
tHi,lh
ti,i,j |
2 +∑
m 6=i,n Ptm,n|d
tHm,nh
tm,i,j |
2 + σ2
)
, ∀t(14)
s.t.∑
j
P ti,j ≤ Pmax, t = 1 . . . T, i = 1 . . .N
Once again, as in the previous case, we assume that the BSs are only allowed
exchange the channel statistics and not the instantaneous CSI values. Hence, the
MBSs optimize the power allocation based on the asymptotic approximations
of the associated quantities which are given by
|dtH
i,jhti,i,j |
2 −
(
σ2i,i,jmΣi
1 + σ2i,i,jmΣi
)2
a.s.−−−−−−→Nt,K→∞
0 (15)
|dtH
i,lhti,i,j |
2 −(σ2
i,i,jσ2i,i,l/Nt)m
′Σi
(
1 + σ2i,i,jmΣi
)2 (
1 + σ2i,i,lmΣi
)2
a.s.−−−−−−→Nt,K→∞
0 (16)
|dtH
m,nhtm,i,j |
2 −(σ2
m,i,jσ2m,m,n/Nt)m
′Σm
(
1 + σ2m,i,jmΣm
)2 (
1 + σ2m,m,nmΣm
)2
a.s.−−−−−−→Nt,K→∞
0 (17)
Here mΣiis the Stieltjes transform of the matrix Σi =
(
∑
m,l hi,m,lhHi,m,l + I
)
.
The reader can refer to [5] for the derivations using random matrix theory. The
details have been omitted in this paper. In order to simplify notations, let us
introduce
Gi,i,j△
=σ2i,i,j
(
1 + σ2i,i,jmΣi
)2 Gi,i,l△
=σ2i,i,l
(
1 + σ2i,i,lmΣi
)2
Gm,i,j△
=σ2m,i,j
(
1 + σ2m,i,jmΣm
)2 Gm,m,n△
=σ2m,m,n
(
1 + σ2m,m,nmΣm
)2
Let us denote the downlink SINR obtained by replacing the numerator and
denominator terms by their asymptotic equivalents by Γi,j ,
Γi,j =Pi,jσ
2i,i,jGi,i,jmΣi
∑
l 6=j Pi,lGi,i,jGi,i,lm′Σi
+∑
m 6=i,n Pm,nGm,i,jGm,m,nm′Σm
+ σ2(18)
15
Note that with the asymptotic approximation, the downlink SINR Γi,j is the
same across the time slots (since Γi,j only depends upon the channel statistics
which is constant across time slots). We re frame the problem with asymptotic
approximations given by
RasympMAX : max
Pi,j
∑
i,j
βi,j log(
1 + Γi,j
)
(19)
s.t.∑
j
Pi,j ≤ Pmax, i = 1 . . .N
The MBSs can now solve the optimization problem RasympMAX by exchanging only
the channel statistics between themselves. We make the high SINR approxima-
tion and use log(1+x) ≈ log(x), for large x. However, the optimization problem
in log(Γi,j) is not concave in terms of P. Hence, we proceed by using geometric
programming approach [20] and make the variable change to Pi,j = log(Pi,j).
RasympMAX : max
Pi,j
log(
exp(Pi,j)σ2i,i,jGi,i,jmΣi
)
(20)
− log
∑
l 6=j
exp(Pi,l)Gi,i,jGi,i,lm′Σi
+∑
m 6=i,n
exp(Pm,n)Gm,i,jGm,m,nm′Σm
+ σ2
s.t.∑
j
exp(Pi,j) ≤ Pmax, i = 1 . . .N
The problem is now a convex optimization problem in terms of the variable
Pi,j (refer to [20] for details). We Introduce the Lagrange variable δi for the
max-power constraint and formulate the Lagrangian as
L(P, δ) =∑
i,j
βi,j log(
exp(Pi,j)σ2i,i,jGi,i,jmΣi
)
(21)
− log
∑
l 6=j
exp(Pi,l)Gi,i,jGi,i,lm′Σi
+∑
m 6=i,n
exp(Pm,n)Gm,i,jGm,m,nm′Σm
+ σ2
(22)
−∑
i
δi
∑
i,j
exp(Pi,j)− Pmax
(23)
Let us also denote the terms inside log in equation (22) by the notation Ii,j . We
also define P =[
P1,1, P1,2, . . . , P1,K , . . . , PN,K
]
. The solution to the optimiza-
16
tion problem can now be given by
minδ
maxP
L(P, δ) (24)
We will use the gradient based method to solve the dual problem [21]. At the
iteration count τ for the gradient based algorithm, for a given δ = δ(τ), let us
define
P(τ) = argmaxP
L(P, δ(τ)) (25)
The update equation for the dual variable is given by
δi(τ + 1) = δi(τ) − α
∑
j
exp(Pi,j(τ)) − Pmax
∀i = 1, . . . , N (26)
In order to solve the optimization problem in (25), we once again use the gradient
based method. The gradient of L(P, δ(τ)) with respect to Pi,j at iteration ζ is
given by
∆i,j(ζ, τ) =∂L(P, δ(τ))
∂Pi,j
∣
∣
∣
Pi,j=Pi,j(ζ,τ)
In particular, we have
∆i,j(ζ, τ) = βi,j −∑
l 6=j
βi,l exp(Pi,l(ζ, τ))Gi,i,lGi,i,jm′Σi
Ii,l
−∑
m 6=i,n
βm,n exp(Pm,n(ζ, τ))Gi,m,nGi,i,jm′Σm
Im,n
− δi(τ) exp(Pi,j(ζ, τ))
We iteratively update the power allocation until convergence
Pi,j(ζ + 1, τ) = Pi,j(ζ, τ) + κ∆i,j(ζ, τ) i = 1, . . . , N, j = 1, . . . ,K
where κ is a small step size. The value Pi,j(τ) is the value of the gradient
algorithm at the convergence point given by,
Pi,j(τ) = limζ→∞
Pi,j(ζ, τ)
17
Now, let use denote the solution of the optimization problem RasympMAX as
µi,j = maxPi,j
RasympMAX
Pi,j = argmaxPi,j
RasympMAX
The MBSs precompute the power allocation based on the channel statistics to
Pi,j and assume a precomputed service rate µi,j . The precomputed service rate
can be viewed as an estimate of the service rate offered to the queues at the CS.
However, in reality, the actual service rate offered to the queues depends on the
channel realization and fluctuates around the target service rate. The service
rate obtained per time slot is given by
µti,j = log
(
1 +Pi,j |d
tHi,jh
ti,i,j |
2
∑
l 6=j Pi,l|dtHi,lh
ti,i,j |
2 +∑
m 6=i,n Pm,n|dtHm,nh
tm,i,j |
2 + σ2
)
The CS has only an estimate of the service rate provided to the UTs given by
µi,j . However, the CS being oblivious to the actual channel realization does not
have the knowledge of the actual service rates in the downlink.
Hence, it is clear that in both the optimization problems (downlink power
minimization and weighted sum rate maximization) there is a mismatch between
the estimated service rate and actual service rate obtained to the queues. We
model the mismatch between the targeted service rate and the actual service
rate as unknown noise and apply H∞ based queue length controller described
in section 4 to regulate the arrival rates into the queues.
In what follows, we provide a flow control algorithm based on H∞ control.
4. Scheduler Design and Queue-Length Stability at the MBS
We now describe the scheduler design. The scheduler has to regulate the
flow into the queues without the knowledge of the wireless channel conditions.
(which implies that the scheduler does not have the knowledge of the actual
received SINR in the downlink, µti,j).
18
Let us now focus on the queue length dynamics at the MBS. The equation
for the queue-length evolution at the MBS is given by
qt+1i,j =
(
qti,j + ati,j − µti,j
)+i = 1 . . .N, j = 1 . . .K (27)
It must be recalled that we assumed the CS to be an infinte reservoir of packets.
The flow controller at the CS (to be described later in this section) specifies ati,j ,
the number of packets that should flow into the queue at the MBS from the CS
at each time slot t and µti,j is the number of packets departing the queue (based
on the downlink SINR corresponding to the beamforming vector design). Let
us subtract the target queue-length qi,j from the LHS and the RHS of (27). We
also add and subtract the target service rate of each queue, µi,j on the RHS of
the equation (27). Hence equation (27) after rearranging becomes
qt+1i,j − qi,j =
(
[qti,j − qi,j ] + [ati,j − µi,j ] + [µi,j − µti,j ])+
(28)
We perform a change of variables and define ψti,j = qti,j − qi,j and the vectorized
version by the notation ψt = [ψt1,1, ψ
t1,2, . . . , ψ
t1,K , . . . , ψ
tN,K ]T . We also define
uti,j = ati,j−µi,j and its vectorized notation by ut = [ut1,1, ut1,2, . . . , u
t1,K , . . . , u
tN,K]T .
Similarly ζt = (µi,j−µti,j) and its vectorized notation ζt = [ζt1,1, ζ
t1,2, . . . , ζ
t1,K , . . . , ζ
tN,K ]T .
Therefore, Equation (28) in its vectorized form becomes
ψt+1 = ψt + ut + ζt (29)
Note that we have removed the (.)+ operation assuming that the queue-length
remains positive at all time instants. The CS makes an observation on the queue-
length in order to regulate the traffic arrival into the queues. The observation
equation is given by
yt = ψt + ηt (30)
where ηt is a noise parameter. The noise parameter is representative of the
fact that the CS only an estimation of queue-length. Moreover, in practical
scenario this observation may even be an outdated observation. We model the
19
observation errors as noise. Note that we do not assume any particular statistical
distribution for the noise parameter.
The objective of the scheduler design problem is to minimize the cost func-
tion given by
J = limT→∞
1
T
[
T∑
t=1
(
||ψt||2W + ||ut||2Q)]
(31)
where W and Q are weight matrices. In our algorithm, both the matrices are
assumed to be identity matrices. The first term of the cost function in Equation
(31) denotes the penalty for deviating from the target queue length. The second
term tries to maintain the arrival rate at each queue as close as possible to the
target service rate γi,j . Note that the scheduler is unaware of the error term
ζt, since it does not have the knowledge of the actual received SINR in the
downlink.
Equation (32) represents a linear dynamic system with the state variable de-
noted by the notation ψt, a control parameter denoted by the notation ut and
an unknown process noise denoted by ζt. In particular, Equation (32) denotes
the evolution of the state variable and (31) the quadratic cost to be minimized.
We use a controller based on H∞ control algorithm [7] to solve the above prob-
lem. The task of the controller is to specify an arrival rate into the queue by
calculating the control parameter ut at each time slot t.
The reason why H∞ filter makes controlling decisions without making any
assumptions of the noise processes is becauseH∞ control assumes the worst case
noise. That is, it tries to minimize the cost function assuming the worst case
noise. For this reason, the H∞ control is called mini-max control. Additionally,
the scheduler is trying to minimize the long term cost function. The problem
is a dynamic programming problem involving stochastic processes i.e., during
each time slot the CS has to take control decisions such that the long term cost
function is minimized (and not the instantaneous cost function). For details of
the H∞ control design please refer to the Appendix A.
We would like to state that in practice, we can add a scaling factor α as a
weight to the control parameter. We define a matrix S = diag( 1√α. . . 1√
α). The
20
equation defining the evolution of the state variable becomes
ψt+1 = ψt + Sut + ζt (32)
And hence, the correspondingly, the weight matrix for the control variable be-
comes V = diag(α . . . α). The parameter α can be varied to adjust the conver-
gence rate of the H∞ control.
5. LQG Control Vs H-Infinity
In this section, we compare the effectiveness of our H-infinity controller in
minimizing the fluctuations of the queue-length as opposed to a standard LQG
controller. It must be noted that standard LQG control assumes that the process
noise is Gaussian in nature and minimizes the expected value of the cost func-
tion. In general, the nature of the fluctuations of the service rate to the queues
around the target service rate would depend on the fluctuations of the wireless
channel which is unknown to the CS. In this section, we assume a Gaussian ap-
proximation to model the fluctuations and perform simulations with standard
LQG control (which is known to be optimal under Gaussian noise). We provide
a brief description of a LQG control in Appendix B.
The simulation results (refer to section 6, Figure 7) show that the H-infinity
flow controller performs better than the LQG controller in terms of keeping
the fluctuations of the queue-length to minimum. The above observation shows
that the Gaussian approximation is not a very accurate representation of the
process noise. Hence, using a robust controller such as H∞ control keeps the
queue-length fluctuations to a lower value and ensures stability.
6. Numerical Results
In this section we provide some numerical results to show the effectiveness
of our algorithm. We consider a hexagonal cellular system with a cluster of 3
cells as shown in Figure 2. Each cell has a MBS which serves only the UTs in
its cell. The number of transmit antennas on each MBS scales with the number
21
of UTs in the cell such that their ratio is a finite constant. In particular, we
assume Nt = K. The MBS are connected to the CS. We assume that the target
queue-length of the buffers at all the UTs to be 20kBs. We also assume the
number of channel uses per coherence interval of the channel (B in equation
(4)) as one-third the target queue-length. The channel realizations are assumed
to be i.i.d. across the coherence intervals.
BS1
BS2
BS3
d1,1,1
d2,1,1
d3,1,1
Figure 2: Example of a network with 3 Cells. The crosses represent the location of the BSs and
the dots represent the location of the UTs randomly scattered inside the cells. The distances
of a UT from the three BSs are also provided.
We consider a distance dependent path loss model in which the UTs are
assumed to be arbitrarily scattered inside each cell. In this case, the path loss
factor from MBS i to UTj,k is given by σ2i,j,k =
(
1di,j,k
)β
where di,j,k is the
distance between MBS i to UTj,k, normalized to the maximum distance within
a cell, and β is the path loss exponent which lies usually in the range from 2
to 5 dependent on the radio environment. We normalize the variance of the
total received noise to σ2 = 1. We also assume that no user terminal is within
a normalized distance of 0.1 from the closest BS.
We perform our simulations for the case of downlink power minimization
(Section 3.1) and assume a target SINR of 9 dB for all the UTs during each
time slot. First let us focus on the case with no max-power constraint on the
22
MBSs (Pmax = ∞). We run our simulations for T = 100 time slots. We first
plot the variation of queue-length values (averaged across the UTs) against the
number of UTs per cell in Figure 3. The bubbles represent queue-length values
at each time slot for a given number of UTs (overlapped on the same point) and
the horizontal line represents the target queue-length.
In order to quantify the variation in the queue-lengths around the target
queue size, we define the Normalized Mean Square Error (NMSE) of the queue-
lengths given by
NMSE =1
NKT
T∑
t=1
∑
i,j
(qti,j − qi,j)2
qi,j
The NMSE for the arrival rates also follows a similar definition. We plot the
NMSE of the queue-lengths against the number of UTs in Figure 4.
It can be seen that in both the Figures 3 and 4 (in this case the curve
corresponding to Pmax = ∞), as the number of UTs increases, the variance of
the queue-length becomes lesser. This is in accordance with our beamforming
algorithm. Recall that as the number of UTs become large, the achieved SINR
in the downlink gets very close to the target SINR for every given channel
realization. Hence, the variation of the queue-length around the target also
reduces.
We also plot the NMSE of the arrival rates into the queues specified by the
scheduler in Figure 5. The NMSE of the arrival rates also reduce as the number
of UTs increase. In fact, when the number of UTs goes to infinity, effectively,
we do not require queue-length control anymore. The arrivals into the queue is
then deterministic equal the the target queue-length at all the time slots.
We now introduce the max-power constraint on each MBS. When the max
power constraint of a particular MBS is not satisfied, we use the approach
developed in Section 3.1.1 to redesign the system parameters. In Figure 4 and
5 (curves corresponding to Pmax = 30, 40), we plot the NMSE of the queue-
lengths and the arrival rates with power constraints. Please note that Pmax
indicated in the plots is normalized by K, the number of UTs. It can be seen
that as the Pmax per MBS reduces, the NMSE parameter increases. This goes
23
well with the intuition since with the max-power constraint, the error between
the target SINR and actual achieved SINR in the downlink increases because
of redesigning of system parameters. The max-power plots will be useful in
determining the excess buffer capacity beyond the target queue-length that the
system designer will have to provide.
0 10 20 30 40 500
0.5
1
1.5
2
2.5
3x 104
No. of UTs per Cell
Que
ue−
Leng
ths
Target Queue−Length
Figure 3: Scatter Plot showing the variation of queue-length around the target, Target SINR
= 9dB, Target Buffer Length = 20kBs, β = 3.6, T =100
5 10 15 20 25 30 35 40 45 500
500
1000
1500
2000
2500
3000
No. of UTs per Cell
NM
SE
Que
ue−
Leng
th
Pmax
= 30
Pmax
= 40
Pmax
= ∞
Figure 4: NMSE of Queue-Length Vs The No. of UTs. Target SINR = 9dB, Target Buffer
Length = 20kBs, β = 3.6, T = 100
In Figure 6, we plot the average downlink transmitted power per UT and the
24
5 10 15 20 25 30 35 40 45 500
500
1000
1500
2000
2500
No. of UTs per Cell
NM
SE
− A
rriv
al R
ate
Pmax
= 30
Pmax
= 40
Pmax
= ∞
Figure 5: NMSE of Arrival Rates Vs The No. of UTs. Target SINR = 9dB, Target Buffer
Length = 20kBs, β = 3.6, T = 100
10 20 30 40 50 60 701
1.5
2x 104
Max−Power Constraint
Avg
. Rat
e A
chie
ved
per
UT
10 20 30 40 50 60 700
50
100
Avg
. Pow
er C
onsu
med
per
UT
Figure 6: Mean Achieved Rate per UT (left y-axis) and Mean Power Consumed (Right y-
axis) per UT Vs Max-Power of the BSs. Target SINR = 9dB, Target Buffer Length = 20kBs,
β = 3.6, T =100, 10 UTs per cell
25
average rate achieved per UT against the given max-power constraint (Pmax) in
the same plot with 10 UTs per cell and 100 channel realizations. It can be seen
that more the power consumed, more is the average rate achieved per UT . This
implies that with more power, the MBSs can meet the targeted SINR without
having to reset them to lower values. Hence, the error between requested SINR
(target) and actual achieved SINR is lower. Also, for low levels of Pmax, it can
be seen that the increase in total downlink transmitted power is almost linear
as a function of Pmax. This is due to the fact that at low Pmax, all the MBSs are
transmitting nearly at their peak in order to provide the best possible data rates
for the UTs (with the power constraint). However, as Pmax increases beyond
certain value, the system performance is close to the case of Pmax = ∞. This
is due to the fact that with higher powers, the targeted SINRs lie inside the
capacity region.
10 20 30 40 500
200
400
600
800
1,000
1,200
No. of UTs per cell
NMSE
Queue-Length
LQG Control
H-Infinity Control
Figure 7: Comparison of H∞ control Vs LQG control, Target SINR = 9dB, Target Buffer
Length = 20kBs, β = 3.6, T = 100
We finally plot the comparison of H∞ control in comparison with LQG
control for the flow controller in Figure 7. The H-infinity controller provides
better performance in terms of minimizing the fluctuations of the queue-length
26
as discusses in Section 5.
7. Conclusions
In this work, we addressed the joint problem of traffic scheduling using H∞
control theory and decentralized beamforming design using tools from random
matrix theory in the context of cellular networks. The task of the H∞ based
flow controller is to regulate the arrival process to the queues at the MBS from
the CS. The H∞ based flow controller stabilizes the queue-lengths at the MBS
without the knowledge of the wireless channel between the MBS and the UTs.
The H∞ controller keeps the fluctuations around the target queue-length to a
minimum as compared to LQG based control.
In this work we decoupled the two problems of traffic scheduling and beam-
forming design. A natural extension of this work would be to jointly address the
two problems of flow control and beamforming design and develop a cross-layer
model to address the same.
Appendix A: H-Infinity Control
In this section, we provide a brief overview on the solution to the discrete-
time system dynamic with unpredictable noise. (Please note that in order to
maintain consistency with the previous literature in this field, we deviate from
the guidelines set for notations in the introduction section this paper). We
consider the following state space equation:
xt+1 = Axt +But +D1Γt (33)
where state vector X, control vector U and noise vector Γ take values respec-
tively in Rn, Rp and Rm. Moreover, we define the following measurement equa-
tion for the state (i.e. the state is not known exactly but estimated via an
observation),
yt = Cxt +D2Γt (34)
27
observable state vector Y takes values in Rn. The objective is to design a
controller that minimizes the following cost function:
L = limT→∞
{ 1
T
T∑
t=0
(
||xt||2W + ||ut||2Q
)}
(35)
where W and Q are the norm matrices. The problem to solve is then a linear
control problem with quadratic cost and unpredictable noise. Please note that
when the noise is Gaussian, the solution to this problem is to use LQG (Linear
Quadratic Gaussian) technique. When the noise is unpredictable (we don’t
have information on the distribution of the noise), an efficient way to solve the
aforementioned problem is to use H∞ technique. The H∞ controller adopted
in this section is the robust controller discussed in [7] and based on zero sum
game theory. To solve our problem using H∞, we introduce the following cost
function:
Jπ = limT→∞
{ 1
T
T∑
t=0
(
||xt||2w + ||ut||2Q − π2||Γt||2)}
(36)
where, π is the level of attenuation. This is a minimax optimization problem
where the cost function is minimized over maximum unknown disturbance. This
problem is a zero sum game with two players. The cost Jπ is minimized by player
1 and maximized by player 2 using vectors U and Γ. One can refer to [7] for more
general class of discrete-time zero-sum games, with various information patterns,
where sufficient conditions for the existence of a saddle point are provided when
the information pattern is perfect and imperfect state.
The H∞ controller can be obtained according to the following proposition
[7];
There exists a state feedback H∞ controller Ut such that,
ut = M(I+ (1−1
π2)Σ)−1x
t (37)
where, ρ(ΣM) < π2. Moreover,
1) M is a minimal non-negative definite solution to the following algebraic
Riccati equation,
M = I+M(I+ (1−1
π2)M)−1 (38)
28
where,
π2I−M > 0 (39)
2) Σ is a minimal non-negative definite solution to the following algebraic Riccati
equation,
Σ = I+Σ(I+ (1−1
π2)Σ)−1 (40)
where,
π2I− Σ > 0 (41)
3) and the state estimate Xtis generated by,
xt = (I+Σ− π−2ΣM)−1(xt +Σyt) (42)
and
xt+1 = xt − ut +Σ(I+ (1−1
π2)Σ)−1(yt − (1−
1
π2)xt) (43)
The above proposition allows us to find the optimal controller for given π.
Therefore, to obtain the solution, we have to find the optimal value of π2 which
satisfies all the above conditions. For that, we use the following lemma: We use
the following algorithm to find the optimal π [ [12], [7], [22] [23] ].
1. Start with a small value of π2 ≥ 0.
2. Increment the value of π2 by a small step size δ.
3. Formulate the two Hamiltonian matrices corresponding to the Ricatti
Equations
H1 = H2 =
I −(
1− 1π2
)
I
−I −I
For our particular problem, the two Hamiltonian matrices H1 and H2 are
identical.
4. Find the eigen values of H1 and H2.
5. If the eigen value of H1 or H2 has an imaginary part, then go to step 2.
ELSE for this value of π2 solve the Ricatti Equations in (38) and (40) to
find the values of M and Σ.
6. If M < 0 or Σ < 0, go to step 2 ELSE check conditions (39) and (41).
29
7. If conditions (39) and (41) are not satisfied, go to step 2. Else check the
spectral radius of ΣM.
8. If ρ(ΣM) < π2, go to step 2, ELSE π2 is the optimal value and terminate
the algorithm.
Appendix B: LQG Control
We now provide a brief description of standard LQG control algorithm. Con-
sider the following state space equation and the observation equations given by
xt+1 = Axt +But +wt
yt = Cxt + vt
where xt is the state variable, ut the control and wt and vt being the process
and the observation noise respectively. The sequences vt and wt are Gaussian
distributed with zero mean and covariance matrix given by
E[
wtT ws]
= R1δt,s
E[
wtT vs]
= R2δt,s
E[
wtT vs]
= 0
where δt,s are Dirac functions. The cost function to be minimized is given by
J = limT→∞
1
TE
[
T−1∑
t=0
||xt||2W + ||ut||2Q
]
The LQG control with imperfect state observations is summarized as follows
[24]. Starting with the initial state estimate xt∣
∣
t=0= 0, solve the following
Ricatti equations given by
S = ATSA−ATSB(
BTSB+Q)−1
BTSA+W
P = A(P −PC(CPCT +R2)−1CP)AT +R1
The optimal control strategy which is a linear control law in the state estimate
is given by
ut = −Lxt
30
where the matrix L is given by
L =(
BTSB+Q)−1
BTSA
The state estimate at time t+ 1 is given as a function of the state estimate at
t and observation yt is given according to the following equation
xt+1 = Axt +But +K(yt −Cxt)
where the matrix K is given as
K = APCT (CPCT +R2)−1
[1] M.-S. Alouini and A. Goldsmith, “Area spectral efficiency of cellular mobile
radio systems,” IEEE Transactions on Vehicular Technology, vol. 48, pp.
1047 –66, 1999.
[2] J. Hoydis, M. Kobayashi, and M. Debbah, “Green Small-Cell Networks,”
IEEE Vehicular Technology Magazine, vol. 6, no. 1, pp. 37 –43, march 2011.
[3] [Online]. Available: http://ecoscells.twn-bl.homeip.net/
[4] [Online]. Available: http://www.willcom-
inc.com/en/service/why/index.html
[5] S. Lakshminarayana, J. Hoydis, M. Debbah, and M. Assaad, “Asymp-
totic Analysis of Distributed Multi-cell Beamforming,” in PIMRC, Istanbul,
Turkey, Sep, 2010.
[6] R. Couillet and M. Debbah, Random Matrix Methods for Wireless Com-
munications. Cambridge University Press, 2011.
[7] T. Basar and P. Bernhard, H-∞ Optimal Control and Relaxed Minimax
Design Problems: A Dynamic Game Approach. Birkhauser, Boston, MA,
1991.
[8] A. Eryilmaz and R. Srikant, “Fair Resource Allocation in Wireless Net-
works using Queue-Length-based Scheduling and Congestion Control,” in
31
24th Annual Joint Conference of the IEEE Computer and Communications
Societies. Proceedings IEEE INFOCOM 2005, vol. 3, march 2005, pp. 1794
– 1803 vol. 3.
[9] M. Neely, E. Modiano, and C.-P. L. ;., “Fairness and Optimal Stochastic
Control for Heterogeneous Networks,” in 24th Annual Joint Conference
of the IEEE Computer and Communications Societies. Proceedings IEEE
INFOCOM 2005, vol. 3, march 2005, pp. 1723 – 1734 vol. 3.
[10] X. Lin, N. Shroff, and R. Srikant, “A tutorial on cross-layer optimization in
wireless networks,” IEEE Journal on Selected Areas in Communications,
vol. 24, no. 8, pp. 1452 –1463, aug. 2006.
[11] E. Altman and T. Basar, “Optimal rate control for high speed telecommuni-
cation networks,” in Proceedings of the 34th IEEE Conference on Decision
and Control, 1995., vol. 2, dec 1995, pp. 1389 –1394 vol.2.
[12] N. Hassan and M. Assaad, “Dynamic Resource Allocation in Multi-service
OFDMA Systems with Dynamic Queue Control,” IEEE Transactions on
Communications, vol. 59, no. 6, pp. 1664–1674, june 2011.
[13] E. Bjornson, R. Zakhour, D. Gesbert, and B. Ottersten, “Cooperative Mul-
ticell Precoding: Rate Region Characterization and Distributed Strategies
With Instantaneous and Statistical CSI,” IEEE Transactions on Signal
Processing, vol. 58, no. 8, pp. 4298 –4310, aug. 2010.
[14] B. L. Ng, J. S. Evans, S. V. Hanly, and D. Aktas, “Distributed Downlink
Beamforming With Cooperative Base Stations,” IEEE Trans. Inf. Theory,
vol. 54, no. 12, pp. 5491–5499, 2008.
[15] R. Zhang and S. Cui, “Cooperative Interference Management With MISO
Beamforming,” IEEE Transactions on Signal Processing,, vol. 58, no. 10,
2010.
32
[16] J. Qiu, R. Zhang, Z.-Q. Luo, and S. Cui, “Optimal Distributed Beamform-
ing for MISO Interference Channels,” IEEE Transactions on Signal Process-
ing, 2010, submitted. [Online]. Available: http://arxiv.org/abs/1011.3890
[17] H. Dahrouj and W. Yu, “Coordinated beamforming for the multi-cell multi-
antenna wireless system,” in 42nd Annual Conference on Information Sci-
ences and Systems, CISS 2008, Princeton, NJ, USA, 19-21.
[18] S. Shi, M. Schubert, and H. Boche, “Rate Optimization for Multiuser
MIMO Systems With Linear Processing,” IEEE Transactions on Signal
Processing, vol. 56, no. 8, pp. 4020 –4030, aug. 2008.
[19] S. Christensen, R. Agarwal, E. Carvalho, and J. Cioffi, “Weighted Sum-
rate Maximization using Weighted MMSE for MIMO-BC Beamforming
Design,” IEEE Transactions on Wireless Communications, vol. 7, no. 12,
pp. 4792 –4799, december 2008.
[20] M. Chiang, Geometric programming for communication systems. Founda-
tions and Trends in Communications and Information Theory, vol. 2, no.
1-2, pp. 1-154, August 2005.
[21] S.Boyd and L. Vandenberghe, Convex Optimization. Cambridge University
Press, 2004.
[22] B. Anderson, “An algebraic solution to the spectral factorization problem,”
IEEE Transactions on Automatic Control, vol. 12, no. 4, aug 1967.
[23] S. Boyd, V. Balakrishnan, and P. Kabamba, “On Computing the H Infinity
Norm of a Transfer Matrix,” Math. Contr. Signals Syst., 1988.
[24] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estimation. Englewood
Cliffs, NJ: Prentice Hall, 2000.
33